The quiet discipline of shipping applied AI.

A BI stack that replaced a €50,000/yr controlling role.

Ten weeks, eight dashboards, twenty-three KPIs, and one Node middleware later, the COO has a number he can put in front of investors.

The engagement began, as these often do, with a spreadsheet that nobody trusted. A German marketing agency of some twenty employees had lost sight of its billable efficiency — the single number on which the margin of the business turned. Four teams, two time-tracking systems, three auth protocols, and a Power BI licence no one had yet cracked open. The question was not whether AI could help; the question was whether any software could render the picture at all.

What followed was, by intention, unspectacular. A 186-hour proof-of-concept, settled as an unpaid pilot on the condition that a paid engagement would follow if the numbers held. They held. Team efficiency climbed from forty-nine percent to sixty-five percent within the first month of deployment — an uplift the COO, in a moment of unusual generosity, attributed to a full-time controlling role worth some €50,000 a year. The engagement is now in its second phase.

“It took me three years and it took you three weeks.”

Five minutes from meeting notes to a fileable Jira backlog.five minutes

An agentic tool built for teams tired of the Monday-morning ticketing tax.

Planning meetings generate transcripts. Jira expects epics, stories, acceptance criteria, and estimates. Between the two lies a tedious hour — longer, if the team has lately been undisciplined. This open-source tool collapses that hour into roughly five minutes, and does so honestly: it parses the notes, drafts the ticket tree, and files the result against the Atlassian REST API in a manner any team lead can audit.

The tool is deliberately small. It does one thing, it does not invent requirements, and it runs either as a CLI or as a scheduled action-hook against a repository's notes folder. The decomposition is done by Claude, the validation by Pydantic, and the filing by the ordinary REST endpoint. No shamanism.

The extraction pipeline behind STOXX-600 climate-disclosure research.

A working paper on climate disclosure required structured paragraphs from six years of European annual reports. I built the scraper.

Not every engagement ends on a balance sheet. Some end in a footnote. A working paper on climate disclosure, authored by researchers at LMU Munich, Bocconi, and IESE under the TRR 266 programme, required financial-statement paragraphs from six years of STOXX Europe 600 annual and audit reports — at a scale where manual coding was untenable. The request, relayed through the Accounting chair at TUM, was for a pipeline that could render the corpus into machine-readable form without losing the provenance any serious reviewer would demand.

The pipeline was built in Python, orchestrated around Claude for section identification, and paired with a reproducibility layer so that any row in the final dataset could be traced back to a page in a source PDF. The paper, SSRN 4763140, cites the work in its acknowledgements — a small line of type, but one that has opened doors since.